AI designs a vaccine for viruses that do not yet exist

The idea sounds like science fiction: A vaccine engineered not to fight a single virus, but to anticipate an entire family of pathogens, including ones that have not yet crossed over into humans. Researchers have now taken an early step towards achieving this goal. This is with the first human trial of an artificial intelligence‑designed “universal” coronavirus vaccine.

In a small phase I study involving 39 healthy volunteers, scientists report that their experimental vaccine was safe and well tolerated, while also provoking immune responses against a spectrum of coronaviruses. These include SARS‑CoV‑2, the virus behind COVID‑19, the earlier SARS virus, and several related bat coronaviruses that have long been watched as potential sources of future pandemics.

What sets this effort apart is not only its breadth, but its origin. The vaccine’s active component is a synthetic antigen. This was designed entirely in silico. Rather than isolating a fragment of a real virus and modifying it, the researchers used artificial intelligence to construct a molecular “best guess” at what all dangerous sarbecoviruses (a specific subgroup within the larger family of coronaviruses, of the Betacoronavirus genus) have in common.

From reactive to predictive vaccines

Conventional vaccine design is inherently reactive. Scientists sequence a virus that is already circulating, identify a target, often a surface protein, and build an immunogen to match it. The approach works, but it is forever playing catch‑up. The seasonal flu vaccine is reformulated every year; COVID‑19 boosters have had to be updated repeatedly as new variants emerge.

The AI‑driven approach aims to invert that paradigm. Instead of focusing on what is circulating now, machine learning models are trained on large datasets of viral genomes drawn from both human infections and animal reservoirs. In the case of coronaviruses, that includes extensive surveillance data from bats and other wildlife.

AI does not “predict viral evolution” in the sense of knowing exactly what a virus will become. Instead, it learns the rules, constraints and probabilities that govern how viruses evolve, then uses those patterns to forecast which mutations are most likely—and most dangerous—next.

The algorithm then identifies conserved regions, the structural or sequence motifs that remain relatively stable across evolution, and designs a synthetic antigen incorporating them. The resulting construct, described as a “super‑antigen”, is not a copy of any single virus. Rather, it is a composite representation of shared viral features.

The hope is that an immune system trained on such an antigen will recognise a broader set of threats, including viruses that have yet to emerge. In effect, the vaccine encodes a statistical model of a viral family, rather than a snapshot of a particular strain. As Jonathan Heeney at the University of Cambridge, who led the work, has argued, this could allow vaccines to become “future‑proof” and no longer locked in a reactive cycle of variant tracking and reformulation.

Designing biology in silico

The trial also marks a conceptual milestone in biomedical engineering. While computational tools have long aided vaccine design, for instance, by helping identify epitopes, the antigen here was designed entirely by AI systems, rather than derived from an existing biological template. This reflects a broader shift taking place across the life sciences: the transition from analysing biology to actively designing it. Advances in machine learning, especially in protein structure prediction and generative modelling, have made it possible to explore vast “sequence spaces” that evolution itself has only partially sampled.

In practical terms, this means researchers can now generate candidate antigens that optimise for properties such as stability, manufacturability and immunogenic breadth—objectives that would be difficult to balance using traditional, intuition‑driven approaches.

The DIOSynVax platform, developed from Cambridge research, applies these principles by combining global genomic surveillance data with computational design tools. Its synthetic vaccine candidates aim to encode consensus features across viral families, effectively distilling evolution’s redundancies into a single immunogen.

The phase I trial was necessarily modest in scale, focusing on safety rather than efficacy. Participants aged 18 to 50 received the vaccine at clinical research facilities in Cambridge and Southampton. No significant adverse effects were reported. Encouragingly, immunological analyses indicated that the vaccine stimulated responses against multiple coronavirus targets, including strains not currently circulating in humans. While detailed neutralisation data remain limited, the findings suggest that the “broad spectrum” concept may be viable.

The vaccine was delivered as a DNA construct using a needle‑free microfluidic jet system, another technological departure from conventional intramuscular injection. Such approaches could simplify mass immunisation in the future, though their scalability remains to be demonstrated.

Despite the promise, substantial hurdles remain. A larger phase II study will be needed to confirm the breadth and durability of immune protection. Crucially, it remains unclear whether responses against distant viruses—such as those found only in bats—translate into meaningful protection in real‑world conditions. There is also a fundamental biological challenge. Viruses evolve, but so too does immune recognition. Even broadly targeted responses can wane, and pathogens may still find evolutionary pathways that evade them. The term “universal vaccine” therefore risks oversimplification; what can realistically be achieved may be better described as “broader and more resilient protection”.

If the approach proves successful, its implications could extend well beyond COVID‑19 and its relatives. The same strategy could, in principle, be applied to other viral families characterised by high mutation rates and zoonotic potential, including influenza viruses and filoviruses such as Ebola.

Influenza, in particular, has long been a target for universal vaccine efforts, yet progress has been slow. AI‑guided antigen design could accelerate this work by identifying conserved features across influenza subtypes and generating optimised immunogens more efficiently than traditional methods.

More broadly, the integration of AI into vaccine development could compress timelines in future outbreaks. During the COVID‑19 pandemic, the interval between viral sequencing and vaccine deployment was measured in months—a remarkable achievement. A predictive, AI‑driven framework could shorten this further, potentially allowing candidate vaccines to be tested before an outbreak begins. Hence, the prospect aligns with a growing emphasis on pandemic preparedness. As ecological disruption and global connectivity increase the likelihood of zoonotic spillover, the ability to pre‑empt emerging pathogens has become a strategic priority.

A new model for preparedness

The work also highlights the importance of data—specifically, the global surveillance networks that provide the raw material for AI models. Without extensive genomic sampling of animal viruses, the algorithm would have far less to learn from.

In this sense, the “intelligence” behind the vaccine is distributed: it reflects not only machine learning techniques, but also decades of fieldwork tracking viruses in remote ecosystems. Maintaining and expanding these datasets will be critical if predictive vaccine design is to reach its full potential.

Finally, there are questions of governance and equity. If vaccines can be designed faster and more flexibly, how will they be regulated? How will global access be ensured? And will such technologies widen or narrow existing disparities in healthcare?

For now, the results represent an early but significant proof of concept. An AI‑designed antigen has been tested in humans and has shown that it can safely engage the immune system in a broad, cross‑reactive way. That does not yet amount to a universal coronavirus vaccine. But it does suggest that the rules of vaccine design are beginning to change. Instead of chasing viruses as they evolve, scientists may be learning to anticipate them, using artificial intelligence not just as a tool for analysis, but as a partner in biological design.